GRAFT-ATHENA: AI That Learns from Past Problems and Discovers New Methods
Autonomous agentic framework reduces exponential search to linear and creates novel numerical solvers.
Scientific discovery involves navigating vast combinatorial spaces of methods and parameters, but existing AI systems treat each problem in isolation, wasting prior experience. GRAFT-ATHENA solves this by modeling decision spaces as factored probabilistic trees (Graph Reduction to Adaptive Factored Trees), where each method becomes a single path. This factorization reduces the parameter footprint from exponential to linear, and the resulting paths embed as unique fingerprints in a metric space. When faced with a new problem, the system queries similar past solutions, accumulating methodological experience across domains.
Tested on canonical physics-informed machine learning (PIML) benchmarks, GRAFT-ATHENA outperformed both human-designed and earlier agentic baselines. It tackled real-world engineering challenges: reconstructing Mach-10 flow over the Apollo Command Module from a 1968 report and recovering shear-thinning blood-cell rheology. Remarkably, the system autonomously proposed regularization constraints for ill-posed inverse problems and discovered a spectral PINN method with exponential convergence. These results demonstrate a path toward autonomous laboratories that grow more capable with every solved problem.
- GRAFT reduces the decision space from exponential to linear using factored probabilistic trees, each method as a single path.
- The system autonomously solved Mach-10 flow reconstruction over Apollo Command Module and blood-cell rheology from sparse data.
- It discovered a new spectral PINN with exponential convergence, showing emergent ability to invent novel numerical methods.
Why It Matters
Enables scientific AI that learns from experience, potentially accelerating discovery across physics, engineering, and materials science.